@Sateluco DMD2 based checkpoints show better speed and much more stable anatomy. But creativity and textures are poor. Hope you will find a way to bring back the textures.

Use RES4LYF for ComfyUI, no need to stick to LCM.

You can use negative weight DMD2 lora and DPM++ or something. Needs some tweaking to find the right spot.

@s1nph73940 Negative DMD2 defeats the whole idea to use DMD2 based checkpoint.

@s1nph73940 Running from the problem is not the way forme... )) I'm already doing the training. Hope I will manage to improve texture in upcoming CinEro ILL V6 R7* release.

@homoludens You complained not me. You need to use DMD lora anyway when using any other non DMD trained lora to get back details, so what?

@s1nph73940 Well, if you are using ComfyUI, you could go as far as to merge DMD2 LoRA at negative weight with checkpoint. While its not most correct way to remove it, it would probably work. Entirely depends on how he makes DMD2 models.

@homoludens Creativity is basically impossible, cause DMD2 is what you can take as "soft distillation" layers. Anything thats distilled by logic simply loses creativity.

DMD2 is in some aspects making all models closer to flow matched than iterative. At least goal is basically the same, just slightly different road.

Most creative model would be something that would require 40-50 steps to make proper pic. Which sadly is something community would probably not like.

@Mescalamba it sounds like what I understood outside of distilation scripts usage experience.
But still.. I want to try and train it for better creativity. Look my latest NoiceAI model. Not super creative but not boring as hell.. it's like a Britney Spears.. I hope

@homoludens You could try instead training model with DMD2 LoRA to train original DMD2 base, but Im not entirely sure its "that easy".

DMD2 LoRA is just extract from SDXL DMD2 checkpoint. Unsure if anyone tried to actually further train/finetune that checkpoint on its own.

@Mescalamba I'am actually doing it right now. E40+ model is released (NoiceAI). Not perfect but good enough to test (Beta).

@homoludens Curious where that will lead. Was it harder to set up than regular model? Or does it require specific training?

@Mescalamba Training is the same. I'm training a big Rank LORA with big diverse dataset with similar technical features but very broad narrative. Like full model fine tuning but just a Rank 128+ LORA. The hardest thing is to wash out the ultimate stability of DMD2 LCM render of the base DMD checkpoint to make it completely broken. Then train to fix it...

@homoludens How about you posting a pic with the metadata intact so we can see what you have done and are complaining about. , If your assertion about dmd2 based checkpoints you believe accurate, why download this checkpoint in the first place if creativity and textures are poor (your words) with dmd2, surely you would expect your results so why the insult

@nationalwolftvzz200 no insult. I posted the critics and @Sateluco responded that he is aware of DMD2 CONs. And also I have my WiP to improve textures and creativity. You can check my latest models with images posted. Metadata is intact. You can drop my images into ComfyUI and see the actual workflow. You can do the same with MoP checkpoints and see how the author of this series and all derivatives with similar renders achieve the textures. They use 2Pass approach. 1st pass with LCM-Exponential gives very smooth image with poor details and minimalistic composition. 2nd pass adds texture, but overall style is very realistic with low creativity. You can do it yourself if you are able to use ComfyUI. You can also do the same with my workflows (also 2pass but with Euler A + DPM++ 2S A samplers for example). No secretes there.

I have my GPUs busy by dataset improvements and training. So, I cannot make a comparison right now. But so far no one author of checkpoints based on DMD2 responded me that there is completely no such issue I am talking about.